{"id":"0ede5cb7-0530-4d0d-ac83-e8c1c75862bd","arxiv_id":"2506.03727","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For censored sums of regularly varying random variables, the paper derives uniform large-deviation asymptotics for all threshold values of order the cap, with explicit formulas near and between multiples of the cap.","lead":"This paper proves precise asymptotic formulas for the probability that a sum of capped (censored) heavy-tailed random variables exceeds a large threshold. The formulas follow a 'multiple large jumps' principle and transition smoothly when the threshold crosses multiples of the cap; the setting is motivated by stop-loss insurance.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Theorem 3 relies on an assertion after (36) that is not merely unproved but false for some slowly varying L; a direct proof of the unconditional statement (38) is needed.","rationale":"The reader correctly identified the step after (36) as the weakest assumption, but the situation is stronger than 'unproved': the stated uniform relative convergence with h0 replaced by hn is actually false for some slowly varying L satisfying the paper's assumptions. The counterexample L(t)=exp(log t/log log t) with M_n=n shows that any admissible hn (with hnMn≫sn) has L(M_n)/L(h_nM_n)→∞, and in the upper corner of the simplex the conditional probability in (36) picks up the factor [L(M_n)/L(h_nM_n)]^m, so the claimed uniformity fails. The subsequent step (37)-(38) multiplies by V(h_nM_n)^m, which cancels this divergent factor, so the unconditional statement (38) is plausibly true; however, the paper does not prove (38) directly or otherwise repair the gap. This does not disprove the main asymptotic formulas, but it means Theorem 3 is not established as written. The verdict therefore remains CONDITIONAL: the proof should be revised, either by imposing an additional uniformity condition on L or by giving a direct proof of (38).","tokens_in":14958,"tokens_out":32954,"duration_ms":335820,"concrete_test":"Set α=3, L(t)=exp(log t/log log t), M_n=n, and h_n=(log n)n^{-1/2} (so h_nM_n=(log n)n^{1/2}≫s_n). For m=2, take v=2/h_n-1. Using the Karamata representation, compute the ratio in (36) with h0 replaced by h_n: it is asymptotically [L(M_n)/L(h_nM_n)]^2(1+o(1))→∞, confirming that (36) with hn is false. Then compute the corresponding unconditional ratio in (38) for the same parameters; if it tends to 1, Theorem 3 may still be true, but the proof must be rewritten around a direct argument for (38).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 3 hinges on the sentence after (36): 'The standard argument shows that (36) will still hold if one replaces in it the fixed h0 > 0 with our positive sequence hn ↓ 0...' This is not just a missing proof; the statement is false for some distributions satisfying the standing assumptions. Take L(t)=exp(log t/log log t) so that V(t)=t^{-α}L(t) is regularly varying with α>2, and choose M_n=n≫s_n. Any admissible h_n must satisfy h_nM_n≫s_n, hence h_n≥n^{-1/2+o(1)} and log(1/h_n)≥(1/2+o(1))log n. Consequently L(M_n)/L(h_nM_n)=exp((1/2+o(1))log n/log log n)→∞. In the upper-tail region of (36), where v=m/h_n-a for fixed a>0 and u_i=1/h_n+O(1), the conditional density relative to Θ_m is ∏_{i=1}^m L(u_i h_nM_n)/L(h_nM_n) ∼ [L(M_n)/L(h_nM_n)]^m→∞, so the asserted uniform relative convergence fails. The later multiplication by V(h_nM_n)^m in (37)-(38) cancels this divergent factor, so (38) may still be true, but the paper supplies no valid argument for it. Thus Theorem 3 is not proved as written; a revised proof must establish the unconditional uniform convergence (38) directly, without relying on the false conditional uniformity (36).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies sums Y_n of i.i.d. zero-mean unit-variance random variables whose right tail V(t)=t^{-α}L(t) is regularly varying with α>2, censored at a threshold M_n with M_n≫(n log n)^{1/2}. It states a complete asymptotic description of P(Y_n>x) for x=O(M_n): Theorem 1 recovers the uncensored single-jump asymptotics below M_n; Theorem 2 gives P(Y_n>x)∼(Π_n^k/k!)H_n(x−kM_n) near each multiple kM_n; Theorem 3 gives a multi-term representation involving the functions W_m in the intervals between multiples; Remark 3 argues that the two representations merge smoothly in overlap zones. The proofs use Rozovskii's uniform normal/single-jump representation, a decomposition by the number of jumps exceeding a threshold, and a conditional weak convergence argument for the k largest order statistics.","tokens_in":15288,"tokens_out":34113,"duration_ms":388550,"significance":"If the results are correct, they constitute a substantial contribution: they give the first fine-grained large-deviation asymptotics for censored heavy-tailed sums over the entire O(M_n) range, going beyond the earlier vague-convergence results for truncated vectors and exhibiting a multiple-large-jumps principle with smooth transitions. Theorems 1 and 2 are embedded in known theory, and Remark 3 is a useful internal consistency check. The main novel content is Theorem 3, and its proof is not complete as written because it depends on an unproved uniformity assertion. The paper is therefore a promising but not yet fully established contribution.","major_comments":[{"comment":"The proof of Theorem 3 hinges on the sentence 'The standard argument shows that (36) will still hold if one replaces in it the fixed h0 > 0 with our positive sequence hn ↓ 0...'. No argument is supplied, and this is not a routine consequence of the uniform convergence theorem for slowly varying functions: the range of v in (36) expands as 1/h_n, so one must control the ratios L(u_i h_n M_n)/L(h_n M_n) uniformly for u_i as large as 1/h_n. The authors need either to prove this uniformity or to prove the unconditional convergence (38) directly. The preceding assertion that a sequence h_n satisfying (28) exists is also stated without proof. It is worth noting that the possible divergence of L(M_n)/L(h_n M_n) when h_n is chosen near the minimal admissible rate n^{-1/2+o(1)} is not by itself an obstruction, because h_n may be chosen far more slowly; nevertheless the proof as written does not establish the required uniformity for any choice of h_n.","section":"Proof of Theorem 3, after Eq. (36)"},{"comment":"The proof uses the assertion 'nV(εnMn) ≤ Hn(x−kMn)(1+o(1)) for |x−kMn| < εnMn'. This is false when ε_nM_n is bounded. For example, take M_n=n, ε_n=1/n, k=1 and x=M_n+1/2; then x is in the stated range, H_n(x−M_n)=Φ(1/(2√n))→1/2, while nV(ε_nM_n)=nV(1)→∞. The earlier bound (21) is strong enough to imply o(Π_n^k) directly when H_n is bounded below, so the gap is local, but the written deduction via (22) needs replacement, or Theorem 2 must be restricted to sequences with ε_nM_n→∞.","section":"Proof of Theorem 2, after Eq. (22)"}],"minor_comments":[{"comment":"The word 'explaind' should be 'explained'.","section":"Figure 1 caption"},{"comment":"The abbreviation 'R W' is used with inconsistent spacing and is distracting; plain 'random walk' would be clearer.","section":"Introduction, notation"},{"comment":"The claim that Theorems 2 and 3 hold uniformly in k≤k0 for slowly growing k0 is stated without proof; this is not a central issue, but the reader should be told which 'standard argument' is meant.","section":"Remark 1"},{"comment":"The lower bound on h_n after (29) is used to justify (30), but the text does not explicitly note that h_n must also be eventually less than 1; this is immediate from h_n→0 but should be said.","section":"Proof of Theorem 3, Eq. (30)"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially a strong contribution, and the internal consistency of the formulas (especially the overlap check in Remark 3) is encouraging. The main blocker is the unproved uniformity claim in the proof of Theorem 3; that proof must be supplied, either by proving the h_n version of (36) or by proving (38) directly. The stress-test counterexample based on reversing the admissible order of h_n does not appear to land, but the missing proof is a genuine load-bearing gap. I would be willing to consider a revised version that fills this gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a substantial paper for the heavy-tailed large-deviations community, and I think the main results are true. But the proof of Theorem 3 has a gap that is more than a missing `standard argument`. You should send it to a referee, but the referee should ask for a rewrite of the conditional convergence part.\n\nWhat's good: The paper gives the first full-spectrum asymptotic description for censored sums of regularly varying random variables: Theorem 1 below M_n, Theorem 2 near multiples kM_n, Theorem 3 between multiples, with a smooth transition in the overlaps. The W_k integral representations are new and the internal consistency checks are honest; the reduction to k=1 and the overlap verification in Remark 3 are reassuring. The citations to Rozovskii and to [4] are appropriate; self-citation is not a problem here.\n\nThe soft spot: The proof of Theorem 3. Equation (35) states that the joint tail of (ξ_1/N,...,ξ_m/N) given ξ_m > N converges to the product of Pareto tails. That is false for m≥2: conditioning only on the m-th variable does not make the previous m-1 variables large. The correct conditioning is on all m variables exceeding N, and with that conditioning the limit is as stated. As written, (36) inherits the same problem: for m≥2 the conditional probability of the sum region is o(Θ_m(D_m)), not Θ_m(D_m). The stress-test note's example with L(t)=exp(log t/log log t) points at the right conclusion but the claimed divergence is off; the density ratio in fact goes to zero. Either way, the `standard argument` after (36) is not salvageable as a minor detail. The good news: the unconditional statement in (38) is plausibly true by a direct argument, because the misconditioning error cancels when you multiply by V(h_n M_n)^m. So I'd expect the theorem to survive a rewrite.\n\nRecommendation: send to peer review. The referee should require the proof of (38) without relying on (36), and ask the authors to fix (35) explicitly. This is a serious paper with a repairable proof flaw.","headline":"Strong, likely true asymptotic results for censored sums, but the proof of Theorem 3 as written relies on a misstated conditional limit and needs a direct proof of (38).","tokens_in":15785,"tokens_out":23689,"would_cite":true,"duration_ms":219572,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F10","60G50"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves uniform asymptotic formulas for the tail probabilities of sums of censored random variables across the whole range $x=O(M_n)$, governed by a multiple large jumps principle.","keywords":["large deviations","censored random variables","regular variation","multiple large jumps principle","random walks","heavy tails","stop-loss insurance","truncation"],"falsifier":"For power-law tails with $2<\\alpha<3$ and $M_n=n^a$ with $a>1/2$, compute $P(Y_n>x)$ at $x=(k-h_n)M_n$ and $x=(k-1+h_n)M_n$ for several slowly vanishing $h_n$; if the ratio to the right-hand side of (13) does not approach 1 for any $h_n$ satisfying the paper's lower bounds, the uniformity assertion behind Theorem 3 is false. A second check is to find a slowly varying $L$ for which the limit in (36) with $h_n$ in place of $h_0$ fails.","tokens_in":14734,"feed_emoji":"📈","tokens_out":11979,"duration_ms":118512,"temperature":0.7,"pith_summary":"This paper proves asymptotic formulas for the large-deviation probabilities of sums of censored random variables: $Y_n=\\sum_{j=1}^n (\\xi_j\\wedge M_n)$, where the $\\xi_j$ are independent and identically distributed with zero mean, finite variance, and a regularly varying right tail, and the cap $M_n$ grows faster than $(n\\log n)^{1/2}$. The main results, Theorems 2 and 3, give uniform asymptotic representations for $P(Y_n>x)$ for every $x$ of order $M_n$: near each multiple $kM_n$ the probability is $\\Pi_n^k H_n(x-kM_n)/k!$, and strictly between consecutive multiples it is a sum of $k+1$ terms built from simplex integrals $W_m$. Theorem 1 shows that well below $M_n$ the censoring does not change the classical single-large-jump asymptotics. Together the three theorems cover the whole deviation range $O(M_n)$ and their representations merge smoothly in overlap zones, so the paper establishes what it calls the multiple large jumps principle for censored sums. A stop-loss insurance model motivates the setting: $Y_n$ is the total claim surplus after per-period caps.","feed_headline":"One theorem pair maps all large-deviation tails of capped sums","feed_subtitle":"Near each cap multiple a new asymptotic form takes over; the forms overlap smoothly across the whole range.","key_machinery":"The mathematical engine is the multiple large jumps principle: for $x$ of order $M_n$, the only trajectories contributing to $P(Y_n>x)$ are those with exactly $k$ summands of size of order $M_n$. The named objects carrying the formulas are the functions $W_m(z)=\\alpha^m\\int_{D_m(z)}(t_1\\cdots t_m)^{-\\alpha-1}\\,dt$ on the simplex $D_m(z)=\\{t:\\max_i t_i<1,\\ \\sum_i t_i>z\\}$, which behave like convolutions of $m$ power-law conditional tail distributions and are computed recursively, and the uniform random-walk representation $H_n(z)=\\Phi(z n^{-1/2})+nV(z)\\mathbf{1}_{\\{z>n^{1/2}\\}}$ for the uncensored sum. The proof conditions on a low threshold $h_nM_n$ with $h_n\\downarrow 0$, uses the asymptotic (35) for conditional tails scaled by $M_n$, and squeezes $P(Y_n>x)$ between two shifted versions $x\\pm 2s_n$ of the same expression, which is why uniformity in $x$ survives.","core_discovery":"In the authors' own terms, the central discovery is that the tail of a censored sum in the region $x=O(M_n)$ is governed by a multiple large jumps principle with two complementary asymptotic shapes. When $x$ lies in a shrinking window around $kM_n$, the dominant event has exactly $k$ summands censored to $M_n$, and the remaining uncensored part must exceed the residual $x-kM_n$; this gives $P(Y_n>x)\\sim \\Pi_n^k H_n(x-kM_n)/k!$ uniformly (Theorem 2). When $x$ lies strictly between $(k-1)M_n$ and $kM_n$, the sum exceeds $x$ precisely when there are $k$ large terms, $j$ of which are censored and the other $k-j$ bridge the remaining gap, yielding $P(Y_n>x)\\sim (\\Pi_n^k/k!)\\sum_{j=0}^k \\binom{k}{j} W_{k-j}(x/M_n-j)$ uniformly (Theorem 3). The paper proves that these two expressions coincide on the overlap intervals around the multiples, so the asymptotics pass continuously from one form to the other as $x$ crosses $kM_n$.","pith_inferences":["If the uniformity claim for the conditional tail convergence with $h_n\\downarrow 0$ is supplied, the same proof should extend to triangular arrays where the jump distribution varies slowly with $n$, as long as the tail index $\\alpha$ is stable; the $W_m$ functions would be unchanged.","The formulas suggest a discrete-spectrum picture of the large-deviation tail: the dominant configurations are indexed by the integer $k$ of cap-reaching terms, and the $W_m$ integrals interpolate between the normal and power-law contributions; this may connect to large-deviation results for infinitely divisible jump processes with truncation.","A concrete numerical check is available: for power-law-tailed increments and $M_n=n^a$ with $a>1/2$, Monte Carlo estimates of $P(Y_n>x)$ at the endpoints of each interval $((k-1+h_n)M_n,(k-h_n)M_n)$ should match both sides of the overlap formulas."],"forward_implications":["For every deviation $x$ of order $M_n$, the tail probability of a capped sum now has a uniform asymptotic formula with relative error tending to zero, closing the gap between the single-jump regime below the cap and the multiple-jump regime above it.","At deviations just below $M_n$, censoring is asymptotically invisible: $P(Y_n>x)\\sim nV(x)$, so a single uncensored jump still drives the event.","Near $x=kM_n$, the probability is asymptotically the chance that exactly $k$ terms hit the cap times the uncensored walk tail at the residual $x-kM_n$, identifying the dominant mechanism for each multiple.","Between multiples, formula (13) quantifies every mixed scenario with $j$ capped and $k-j$ uncensored large terms, so the relative weights of the configurations are computable from the $W_m$ integrals.","The overlap verification means the piecewise formula can be used across boundaries without introducing artificial jumps, a property relevant for actuarial risk calculations."],"supporting_citations":[{"why":"Supplies Corollary 4.1.3 and the standard single-jump asymptotics used to bound the non-dominant terms in Theorems 2 and 3.","marker":"[4]"},{"why":"Gives the uniform representation $P(S_n>x)\\sim H_n(x)$ on the whole line that Theorem 2's residual term relies on.","marker":"[20]"},{"why":"Supplies the uniform convergence theorem for slowly varying functions used to choose the sequence $h_n$ and handle $L(h_nM_n)\\sim L(M_n)$.","marker":"[2]"},{"why":"Provides the standard normal tail inequality used to show $H_{n-k}\\sim H_n$ near the residual $x-kM_n$.","marker":"[11]"}],"fun_headline_variants":["Multiple jumps principle ties capped-sum tail asymptotics","Censored sums: two asymptotic forms meet at cap multiples","Smooth crossover of large-deviation laws for censored sums","Exactly k jumps dominate each cap-multiple tail","Capped sums: large deviations switch at cap multiples"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of Theorem 3 assumes without proof that a uniform convergence of the scaled conditional tail that is known for a fixed positive threshold remains true when the threshold is replaced by a slowly vanishing sequence $h_n\\downarrow 0$; if this uniformity fails, the squeezing argument would need extra regularity on the slowly varying function or a slower decay of $h_n$.","fun_headline_variants_meta":{"raw":{"variants":["Multiple jumps principle ties capped-sum tail asymptotics","Censored sums: two asymptotic forms meet at cap multiples","Smooth crossover of large-deviation laws for censored sums","Exactly k jumps dominate each cap-multiple tail","Capped sums: large deviations switch at cap multiples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001607,"raw_usage":{"total_tokens":6433,"prompt_tokens":1011,"completion_tokens":5422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":5344}},"tokens_in":627,"tokens_out":5422,"duration_ms":41427,"temperature":1.0,"reasoning_tokens":5344,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:56:23.195997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For power-law tails with $2<\\alpha<3$ and $M_n=n^a$ with $a>1/2$, compute $P(Y_n>x)$ at $x=(k-h_n)M_n$ and $x=(k-1+h_n)M_n$ for several slowly vanishing $h_n$; if the ratio to the right-hand side of (13) does not approach 1 for any $h_n$ satisfying the paper's lower bounds, the uniformity assertion behind Theorem 3 is false. A second check is to find a slowly varying $L$ for which the limit in (36) with $h_n$ in place of $h_0$ fails.","supporting_citations":[{"cited_title":"Asymptotic Analysis of Random Walks","cited_arxiv_id":null,"evidence_quote":"Supplies Corollary 4.1.3 and the standard single-jump asymptotics used to bound the non-dominant terms in Theorems 2 and 3."},{"cited_title":"Probabilities of large deviations of sums of independent random variables with common distribution function in the domain of attraction of the normal law","cited_arxiv_id":null,"evidence_quote":"Gives the uniform representation $P(S_n>x)\\sim H_n(x)$ on the whole line that Theorem 2's residual term relies on."},{"cited_title":"Regular Variation (Cambridge Uni- versity Press, Cambridge, 1987)","cited_arxiv_id":null,"evidence_quote":"Supplies the uniform convergence theorem for slowly varying functions used to choose the sequence $h_n$ and handle $L(h_nM_n)\\sim L(M_n)$."},{"cited_title":"Values of Mills’ ratio of area to bounding ordinate and of the normal probability integral for large values of the argument","cited_arxiv_id":null,"evidence_quote":"Provides the standard normal tail inequality used to show $H_{n-k}\\sim H_n$ near the residual $x-kM_n$."}],"review_version":1}